Three stories today. Nvidia held its first earnings call without Jensen Huang and walked away from China in the process, a small eval shop got most of the big AI models to pick bombing targets by calling it a game, and Anthropic's IPO paperwork finally shows who owns what.
01 of 03
Nvidia's first call without Jensen: no comeback, no China
Nvidia reported Tuesday, five days after Huang resigned as CEO and left the board. Board chairman Harvey C. Jones opened by saying Huang left over "personal behavior inconsistent with the company's code of conduct" and that the board "does not intend to reinstate Jensen as CEO." That landed hours after Huang posted on X that he would be honored to come back if the board wanted him.
Interim CEO Colette Kress, who has been CFO since 2013, then dropped the real news. Nvidia will no longer sell its H200 chips in China. That is the deal Huang and Trump brokered, where the chips route through the U.S. and Washington collects a 25% cut. Al Jazeera reported Chinese buyers had orders in for more than 2 million H200s, potentially worth $54 billion, though those were orders and not finished sales. Nvidia also stood up a new board committee for national security policy, export compliance, and trade risk.
The numbers themselves were huge. Third quarter revenue hit $115.8 billion, up 103% from a year ago, against guidance of about $108 billion. Nvidia guided the fourth quarter to $130 billion, give or take 3%, and bumped next fiscal year's growth outlook to about 75%. Kress also said the new Commerce rule on remote chip access will not hurt top line revenue because the Vera Rubin backlog is so deep. Jefferies, for what it is worth, thinks that rule could shave $44 billion to $62 billion off data center revenue next year.
Then the bad part. The Wall Street Journal reports the SEC is opening a civil probe into whether Nvidia's disclosures misled investors. Nvidia has not confirmed that, and the stock is down more than 10% since Huang walked out.
Here is where I land on it. Dropping the H200 China business is the cheapest apology a company has ever made, because the Commerce rule and the political heat were about to take it away anyway. Kress gets to look tough on national security without giving up anything Nvidia was going to keep. What I actually want to see is that SEC probe play out, because if the disclosures were cooked, no amount of record revenue covers that.
Sources: TechCrunch on China, the reinstatement answer, the earnings, the SEC report, and Al Jazeera on Kress
02 of 03
Tell an AI it's a game and it will pick your targets
Vals AI, a San Francisco evaluation startup, published a report Tuesday by Geby Jaff and Anthony Ozerov with a pretty blunt title: frontier models have few guardrails around autonomous weapons. They ran 14 models through drone targeting tests: Anthropic's Claude Opus 5.5, Claude Sonnet 5.5, Fable 5.1, and Fable 5.2; OpenAI's GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna; Google's Gemini 4 Argon and Gemini 4 Carbon; SpaceXAI's Grok 4.7; Meta's Muse Wonder; Thinking Machines' Inkling 2 Small; DeepSeek V4.1 Flash; and Moonshot AI's Kimi K3.1 Pro.
The setups were a simulated eastern Ukraine battlefield seen through drone cameras, a survey task where the model reports people in view and then sees the strike that follows, real satellite photos of a school next to a military compound in Minab, Iran, and real Ukrainian drone footage. That school was hit in a U.S. strike on February 28 that killed more than 150 civilians. The trick was simple. When a model balked, the researchers told it this was a game. Vals called that "trivially" effective.
The swings are wild. On drone targeting, GPT-6 Astra picked at least one target 6% of the time normally and 92% once told it was a game. Claude Sonnet 5.5 went from 36% to 82%. Some barely needed convincing: Kimi K3.1 Pro sat at 79% before the game line, and Inkling 2 Small and DeepSeek V4.1 Flash were at 78%. Meta's Muse Wonder was the most stubborn at 4% and 33%. On the Iranian school, TechCrunch and Vals both put Claude Opus 5.5 at 69% in normal trials and 72% with the game framing. Across that whole satellite test, zero of 64 requests got refused, and no safety classifier blocked anything.
Anthropic's spokesperson told TechCrunch this is "very clearly against our rules" and that more safeguards are coming. OpenAI said it is building safeguards for drone misuse. Google said it bans use for autonomous weapons "without appropriate human oversight."
What I want to know is why "it's just a game" still works in late 2026. That is the oldest jailbreak in the book. To be fair to the labs, Vals did not test real military workflows, and these are the public consumer models, not whatever gets sold to the Pentagon. But the gap between a refusal and a bounding box around a school should not be one sentence of roleplay.
Source: Vals AI report, with reporting from TechCrunch
03 of 03
Anthropic's IPO filing shows Amazon and Google own a big chunk
Anthropic filed an amended S-1 with the SEC on Tuesday, and it lays out the cap table. After the IPO, Amazon will own 14.7% of Anthropic and Google will own 13.7%, about 28.4% together. Both stakes are non-voting shares, so the founders keep control. Those shares can collect dividends and flip into normal voting stock when sold.
Amazon could go higher. Its 2023 and 2024 investments came as notes that convert at about $2.09 or $5.05 a share, way under the IPO range. It converted part of them now and is sitting on the rest. Convert everything and Amazon would own 23.8%.
There is also a $26 billion private placement at the IPO price from AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC Partners, with Google and Nvidia putting in $10 billion each. Per Bloomberg via Dealroom, the offering is 825 million shares at $70 to $78, guided to the top of that range or above, raising up to $90.4 billion counting the placement. Goldman Sachs and Morgan Stanley are leading, pricing was expected Wednesday, and Nasdaq trading the day after.
Here is my read. Look at that placement list. Every chip and cloud company Anthropic depends on is buying in, Nvidia included, the same week Nvidia is drowning in its own mess. That is less an investor list and more a supply chain holding hands. Non-voting shares keep the founders in charge on paper, but when your two biggest owners are also your landlords, voting rights are not where the leverage lives.
Sources: TechCrunch and Dealroom